Is there a difference ?

What isn’t different - Tasks

• Machine Learning
  • Supervised Learning
  • Classification
  • Unsupervised Learning
  • Density Estimation
• In statistics we call them:
  • Regression
  • Binomial Regression
  • ???
  • Density Estimation

What isn’t (really) different - Tasks

What isn’t different - Techniques

• Techniques:
  • Random Forests
  • Elastic Nets
  • Neural Networks
• Statistics cover the same topics
  • Elements of Statistical Learning
  • Advanced Data Analysis
• Machine Learning courses often start with:
  • Linear Regressions
  • Logistic Regressions

What isn’t (really) different - Techniques (2)

What isn’t (really) different - Models

What is different – Approach

• Objective:
  • Inference
  • Prediction
• Target:
  • Estimate Functional Form
  • Minimize Cross-Validation Error
• Why?
  • If you have large amount of data, you can let it guide the shape of your regression
  • Try it out with a 10th power regression!

Cross-Validation

Freedom (from p-values) isn’t free

Machine Learning, the hype versus reality

Econometric way of dealing with data

• Economist tell you that \(Y \sim X \beta\)
• Gather data on \(Y\) and \(X\)
• Run a regression on them
• Check \(p\) values
• if \(p<=.05\):
  • publish a paper

Machine Learning way of dealing with data - tired

• Domain expert tells you \(X\) affects \(Y\)
• Run non-parametric regression \(Y \sim X\)
• Find \(f(X)\) that best predict \(Y\)
• Good for:
  • Finding deep correlations
  • Use them
• Not good for:
  • settling scores amongst economists
• How’s it different from just throwing a lot of \(X\) at a linear regression?

Machine Learning way of dealing with data - wired

• Start with a very basic set of features \(X\) that may explain \(Y\) once combined
• Gather lots of data
• Find \(f(X)\) that best predict \(Y\)
• If you have truly large sets of data you may be able to avoid needing a domain expert at all

Digits Example

Regression and Pacman (tuning)

Regression and Pacman (tuning)

• Regression
  • Come up with some features \(X\)
  • Collect games data from people playing
  • (logistic) regression:
    \[ \text{Action} \sim f(X) \]
• Application
  • Computer pacman follows regression blindly

Limits to regression alone (cases not observed)

Limits to regression alone (features)

Simple Solution

Difficult Search

\[ \text{Probability of Winning} \sim f(X) \]

Deep Blue

• 8000 features
• Hand tuned

Self Play

• Start with basic features
• Use complicated neural network to produce: \[ \text{Probability of Winning} \sim f(X) \]
• Make computer use that \(f\) to play against itself
• Produce more data, feed back into neural network

Alpha Go

• In a simplified format
  • Start with 10 basic features of go
  • Regress $ f(X)$ on database of human games
  • Self play improvement
  • Beat humans

Alpha Zero

• In a simplified format:
  • like above but with no starting features

It’s really mostly about features

• Neural network does two things:
  • Creates complex features out of base input features
  • Combines them non-linearly
• Only the first one really matters

Summary

• Two ways to generate “features”
  • Expert Knowledge
  • Fancy Neural Networks
• Which one is better?
  • Domain Knowledge vs Architectural Knowledge

Which one is fake?

Overfitting

Data Quality

Black Box Cargo Cult

[Estimating the value of clean air], however, is a methodological rather than a conceptual challenge, and sophisticated statistical techniques have been developed to isolate the effect of air quality from other (potentially correlated) factors.

Extrapolation

Operationalization

• Dell & Querubin (2018)’s Vietnam:
  • Airforce
  • Bayesian Predictor
  • Score from 0 to 5
  • Rounding

Feedback Loops

• Build statistical model:
  • $ () f(X)
  • Use old data
• Use statistical model to target sales
  • Focus on advertising when \(\text{Probability}\) is high
• Use new data to improve statistical model
• Exploration vs Exploitation